Learnable Blur Kernel for Single-Image Defocus Deblurring in the Wild

Recent research showed that the dual-pixel sensor has made great progress in defocus map estimation and image deblurring. However, extracting real-time views is troublesome complex algorithm deployment. Moreover, deblurred generated by deblurring network lacks high-frequency details, which unsatisfactory human perception. To overcome this issue, we propose a novel method uses guidance of to implement The proposed consists learnable blur kernel estimate map, an unsupervised method, single-image generative adversarial (DefocusGAN) for first time. can learn different regions recover realistic details. We loss guide training process. Competitive experimental results confirm with kernel, achieve comparable supervised methods. In task, achieves state-of-the-art results, especially significant improvements perceptual quality, where PSNR reaches 25.56 dB LPIPS 0.111.